Clear Thinking beats deep thinking in LLM models
We built our own LLM model - Class X from scratch having the core focus of enabling clear thinking inside LLMs as opposed to the conventional wisdom of deep thinking prioritised by AI labs. First we made architectural optimisations - 1.New CLEU optimiser 2. Improved data flow between layers 3.Custom CUDA kernels for backward pass What we did to achieve this was, we curated high quality data without nested arguements along with modification of answers to simple problems but which are verbose but not high quality. Overall, curation and annotation of data was the most crucial step in making the LLM think clearly, you can try the model at clearthinkai.com. Would be good to have you guys review the LLM model.
4 comments
Frankly, I've never really felt like "deep thinking" gave me more value from an LLM. Usually seems to get more roundabout information and take longer.